International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 03 | Mar 2019
p-ISSN: 2395-0072
www.irjet.net
Prediction of Traffic Signs for Automated Vehicles using Convolutional
Neural Networks
S. Saranya1, R. Sindhu Sri2, I. Subhashini3, Dr. K. Anitha4
1S.Saranya
Panruti
Sri Chennai
3I.Subhashini Kovilpatti
4Dr.K.Anitha, Associate Professor, Department of Computer Science and Engineering, R.M.K Engineering College,
Tamil Nadu, India.
---------------------------------------------------------------------***--------------------------------------------------------------------2.R.Sindhu
Abstract - Road Traffic accidents is one of the major reasons
for deaths taking place in India. These accidents not only
result into serious injuries but may also lead to deaths. Image
recognition technology is one of the widely used techniques
used in various fields in research like agriculture, medicine,
automobile etc. Image recognition techniques are being used
which are not only time consuming but also very simple to
handle. It also provides improved algorithms, and makes them
more and more efficient and robust. Principles of convolution
neural network are used. Its numerous applications in the
domain of Image Processing helps to handle traffic sign
recognition systems by indicating driverless vehicles about the
approaching traffic signs and gives alerts to the vehicle.
Finally, the challenges faced by Convolution Neural Network in
terms of time complexity and accuracy were analyzed and is
being enhanced by uses more accurate filters and using tensor
flow algorithms for giving accurate result of traffic sign
worsening due to over population in urban cities. Some of
the road traffic sign information can also be obtained from
GPS, but it is not always up to-date. After extraction of the
road traffic signs from the system, they can be displayed on
the panel of the cars, or could injuries or deaths. Traffic
Symbol recognition is basically a methodology in which the
vehicles are able to determine road traffic signs and avoid
the accidents taking place. These signs may include various
road symbol alerts like “speed limits”, “one-ways”, “school
ahead”, etc. Considering the current Traffic management
system, there is high possibility that the driver would miss
out the road traffic symbol plate alerts due to overcrowding
of the traffic on-road. The condition is even worsening due to
over population in urban cities. Some of the road traffic sign
information can also be obtained from GPS, but it is not
always up to-date. After extraction of the road traffic signs
from the system, they can be displayed on the panel of the
cars, or could driving as well.
Key Words: Image Processing, Sobel Edge detection
algorithm, Frame extraction, Frame normalization,
Tensor flow, Convolution Neural Network, Median filter,
Trained Datasets
1.1Literature Survey
Initially image segmentation techniques were used. They
were inaccurate and incompatible with changing
environmental
conditions(1989).Thus
later
in
2016,successful implementation of the system was done
using canny edge detection algorithm. It highlights edges to
compare shape of traffic sign with data set. In (H. Fleyeh et
al, 2003) Image segmentation technique is used to determine
the shape of the traffic sign by comparing it with the
provided data set. Image segmentation technique for road
traffic sign recognition results in inaccurate output due to
blur images captured during motion of the car. In (V. Andrey
et al, 2006), RGB color segmentation technique is used along
with rule based approach. Image morphological analysis is
done to determine the shape of traffic sign. Results obtained
contain large amount of false positives. In (H. GómezMoreno et al, 2010) basically, the approach is based on
Maximally Stable
1. INTRODUCTION
Traffic accidents are a major reason for death and injuries in
the country. The condition is improving in many parts of the
world, whereas large number of accidents is still taking place
in India. A recently conducted survey shows 64% of drivers
in India on an average have regretted missing road traffic
This may also result in road traffic accidents causing into
serious injuries or even deaths at times. Almost 24% of
drivers in India follow Google maps application without even
paying any attention to the on road traffic sign boards (Y.
Hatolkar et al, 2017). This could possibly be another reason
for on road accidents resulting into serious injuries or
deaths. Traffic Symbol recognition is basically a methodology
in which the vehicles are able to determine road traffic signs
and avoid the accidents taking place. These signs may
include various road symbol alerts like “speed limits”, “oneways”, “school ahead”, etc. Considering the current Traffic
management system, there is high possibility that the driver
would miss out the road traffic symbol plate alerts due to
overcrowding of the traffic on-road. The condition is even
© 2019, IRJET
|
Impact Factor value: 7.211
External Regions (MSERs). Support Vector Machine (SVM) is
cascaded for training system. This detection technique is
significantly insensitive to variations in illumination and
lighting condition.
|
ISO 9001:2008 Certified Journal
|
Page 6647
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 03 | Mar 2019
p-ISSN: 2395-0072
www.irjet.net
processed image frames are sent to image detection module
for further processing technique. In binarization, the traffic
symbol is represented by white pixels, whereas the
background is represented by black pixels. This helps to find
approximate location of the road traffic symbol in the image.
The result is obtained in the form of (x,y) coordinates.
1.2 Existing System
Images captured from the camera are usually of poor quality.
The system needs to enhance the quality of the images, in
order to obtain the correct outcome. There are various preprocessing techniques applied before actual classification of
the image. Image is transformed into various color spaces.
Different color spaces such as HSI (A. de la Escalera et al,
2003), Improved HLS (H. Fleyeh et al, 2004), normalized
color space (W. Ritter et al, 1995) and (J. Greenhalgh et al,
2012) are used. Basically, image normalization technique is
used for adjusting the brightness of the image, so as to detect
the traffic symbols accurately. In (Y. Wu et al, 2013) and (M.
Liang et al, 2013), SVM classifier was used in order to train
the system and map each pixel of the frame to gray scale.
Traditionally, threshold based methods (M. Young et al,
1989), and (A.de la Escalera et al, 2003) were used for object
classification using various image segmentation techniques,
and compared in order to get the optimal technique. These
techniques proved less accurate for environmental changes
such as lightening and bright images during sunny days.
Frame extraction is primary step from the videos captured
from car-mounted camera. RGB is an additive color
component model in which red, green and blue colors are
combined in order to get various shades of colors. RBG color
component is used in order to reduce the time and space
complexity of the system. Frame normalization technique is
used to adjust the brightness in the images due to
environmental factors. This was considered as a drawback in
previous papers (W. Ritter et al, 1995), (J.Greenhalgh et al,
2012). In Binarization technique, the traffic signs are
highlight to white pixels, while background remains blurred.
Initially, techniques used for classification involved feature
extraction and classifier training. SVM classifier was used
along with HOG features (I. M.Creusen et al, 2010), MLP
(Multi-Layer Perceptron) having radial features (Y.Jiang et al,
2011), ANN (artificial Neural Network) along with RIBP
(Rotation Invariant Binary Pattern) (S. Yin et al, 2015). In
short, an optimal feature extraction, and accurate classifier is
a challenging job. Application of Gaussian Blur algorithm for
noise reduction .Finding intensity Gradient of the image:
Calculation of intensity gradients G = (Gx2 + Gy2)1/2 θ = tan1(Gy/Gx) Non-maximal suppression: Threshold is chosen so
as to suppress the noise and determine the exact edges.
Further, it is necessary to suppress the non-maximal pixels
for accurate edges detection. Hysteresis Thresholding: edge
> threshmax : Included edge <threshmin : Excluded
threshmax > edge >threshmin : Included (Only if edge is
connected to strongly connected edge).
Shape classification:
Further, after locating the approximate coordinates of the
road traffic from the image frame, Canny edge detection
algorithm is used to highlight the edges of the road traffic
symbol, whose location is obtained from the previous phase.
This output is sent to Convolution Neural network, where
matching of the actual road traffic symbol from the image
frame, and provided template of road sign from the data set
is done. This output is further provided to Fuzzy
classification in order to improve the accuracy of the system.
Fuzzy Classification:
With reference to the paper (H. Luo et al,2016), Fuzzy
classification technique is applied as extension in order to
improve the efficiency of the system and optimize the
outcome of the application. According to the proposed
system, Fuzzy Classification output will be divided into
certain set of probabilities and final result will be extracted
from matching probable class and will in turn be converted
into audio.
2.1 Proposed System
Convolution Neural Network (CNN), on the other hand can
be considered as one of the popular techniques, for training
and classification. In (G. Wang et al, 2013) modified version
of cross entropy loss was used for CNN training. In spite of
the fact that CNN has shown excellent performance in image
classification, the task of designing a good architecture and
train a workable model is still one of the challenging tasks. In
order to handle different geometry variations of road traffic
signs, data augmentation technique was used to enlarge the
training data set(P. Sermanet et al, 2011) and (J. Jin et
al,2014). A video camera is fitted to the vehicle which had
digital sensors to sense the nearby traffic sign boards.
IMAGE PROCESSING:
Image is converted to RGB colour component model. It has
normalized frames to optimize images .Edge detection is
done using sobel operator Their greatest advantage is it’s
simple to handle and provides greater approximation. Its
sensitivity to noise can be overcome using MEDIAN filter
which removes noise from images. Image detection is carried
out to detect location of traffic signals which examines
intensity of each pixel by calculating the threshold for it. It
splits images into grid of cells and treats each as separate
image. It passes through sign detection where various signs
are analysed using tensor flow algorithms .Its then passed
Pre-processing:
In pre-processing phase, image is converted into RGB
color component model. Further, frames are normalized in
order to optimize the image appearance and control excess
brightness due to sunny weather at times. Gaussian filter is
used in order to smooth the image and remove unwanted
noise which can lead to inaccurate results. This pre
© 2019, IRJET
|
Impact Factor value: 7.211
|
ISO 9001:2008 Certified Journal
|
Page 6648
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 03 | Mar 2019
p-ISSN: 2395-0072
www.irjet.net
through the convolutional neural network which is a
classifier where the probabilities for signs are calculated by
providing the trained database as input to it and after a
series of refinements using refinements using tensor flow
algorithms suitable audio signals are generated from the
vehicle and is intimated about the traffic symbol
approaching.
The proposed system reduces time complexity and it
requires little processing through tensor flow algorithm.
CNN requires less number of parameters. Canny edge
detection algorithms are being replaced by sobel algorithm
which is more simple and provides greater approximation in
identifying edges. Median filters remove noise providing
better normalization than Gaussian filter. Through the
pooling layer of CNN larger images can also be easily
handled with fewer parameters. Also threshold values for
pixels can be identified with ease using OTSUs method to
classify black and white pixels by splitting the image into
smaller portions. Classification is performed by probabilistic
controller by its flexible nature. It can sense images
remotely. It improves efficiency and can be used for complex
systems.
3. CONCLUSION
In this paper, we propose a new modified approach for road
traffic sign recognition technique. Approximate position of
the traffic sign is determined, and sent to convolution neural
network for training and classification. It is an optimized
module for improving the results obtained by CNN. This
traffic sign recognition system will help drivers to track the
traffic symbols with ease, and avoid the accidents and in turn
reduce the number of deaths.
Tensor Flow is an open-source software library for dataflow
and differentiable programming across a range of tasks. It is
a symbolic math library, and is also used for machine
learning applications such as neural networks. TPU is a
programmable AI accelerator designed to provide high
throughput of low-precision arithmetic (e.g., 8-bit), and
oriented toward using or running models rather than
training them. Google announced they had been running
TPUs inside their data centers for more than a year, and had
found them to deliver an order of magnitude better
optimized performance per watt for machine learning. It
runs on python. It has flexible architecture providing easy
deployment of computation. The computations are
represented as dataflow graphs.
REFERENCES
[1] Sanjay Kumar Singh (2016), Road Traffic Accidents in
India: Challenges and Issues, Transportation Research
Procedia, 25 (9),pp. 4708-4719.
[2] Hengliang Luo, Yi Yang, Bei Tong, Fuchao Wu, and Bin
Fan (2016), Traffic Sign Recognition Using a MultiTask
Convolutional Neural Network, Intelligent Transportation
System., 4(3), pp. 237–248.
[3] Hasan Fleyeh (2003), Color detection and segmentation
for road and Traffic signs, Cybernetics and Intelligent
System., 21(3), pp. 247– 258. Vavilin Andrey and Kang Hyun
Jo (2006), Automatic Detection and Recognition of Traffic
Signs using Geometric Structure Analysis, SICEICASE, 4(5),
pp. 222–235.
2.2 Results and discussion:
[4] H. Gómez-Moreno, Maldonado-Bascón, Pedro GilJiménez, Sergio LafuenteArroyo (2010), Goal evaluation of
segmentation algorithms for traffic sign recognition,
Intelligent transportation systems, 11(4), pp. 1-7.
[5] A. de la Escalera, J. M. Armingol, and M. Mata (2003),
Traffic sign recognition and analysis for intelligent vehicles,
Image Vis. Comput., 21(3), pp. 247–258.
[6] H. Fleyeh (2004), Color detection and segmentation for
road and traffic signs, IEEE Conf. Cybern. Intell. Syst., 2(4).
Dec. 2004, pp. 809–814.
© 2019, IRJET
|
Impact Factor value: 7.211
|
ISO 9001:2008 Certified Journal
|
Page 6649
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 03 | Mar 2019
p-ISSN: 2395-0072
www.irjet.net
[7] W. Ritter, F. Stein, and R. Janssen (1995), Traffic sign
recognition using color information, Math. Comput. Model.,
22(5), pp. 149–161.
[8] J. Greenhalgh and M. Mirmehdi (2012), Real-time
detection and recognition of road traffic signs, IEEE Trans.
Intell. Transp. Syst., 13(4), pp. 1498–1506, Dec.
[9] Y. Wu, Y. Liu, J. Li, H. Liu, and X. Hu (2013), Traffic sign
detection based on convolutional neural networks, in Proc.
Int. Joint Conf. Neural Netw. (IJCNN), 4(7), pp. 1–7.
[10] M. Liang, M. Yuan, X. Hu, J. Li, and H. Liu (2013), Traffic
sign detection by ROI extraction and histogram featuresbased recognition,” in Proc. Int. Joint Conf. Neural Netw.
(IJCNN), 3(9), pp. 1–8. M. Young (1989), The Technical
Writer’s Handbook. Mill Valley, CA: University Science 9(2),
pp. 88-98.
[11] I. M. Creusen, R. G. J. Wijnhoven, E. Herbschleb, and P. H.
N. de With (2010), Color exploitation in hog-based traffic
sign detection, in Proc. 17th IEEE Int. Conf. Image Process.
(ICIP), 4(8), pp. 2669– 2672.
[12] Y. Jiang, S. Zhou, Y. Jiang, J. Gong, G. Xiong, and H. Chen
(2011), Traffic sign recognition using ridge regression and
OTSU method, in Proc. IEEE Intell. Veh. Symp. (IV), 9(8), pp.
613–618.
S. Yin, P. Ouyang, L. Liu, Y. Guo, and S. Wei (2015), Fast traffic
sign recognition with a rotation invariant binary pattern
based feature, Sensors, 15(1), pp. 2161– 2180.
© 2019, IRJET
|
Impact Factor value: 7.211
|
ISO 9001:2008 Certified Journal
|
Page 6650